Objective: Continuous glucose monitoring (CGM) provides a novel approach to monitor postprandial glucose responses (PPGR) in persons with gestational diabetes mellitus (GDM). We sought to develop a machine learning (ML) framework to identify PPGRs automatically from CGM data to address challenges of manual meal logging.
Methods: Adults 18+ years diagnosed with GDM were enrolled at 24-32 gestational weeks and wore a blinded Freestyle Libre CGM and logged meals for up to 14 days each over 3 visits. A random forest ML algorithm was trained to identify the first meal of the day from raw daily CGM profiles. Comparing self-recorded and ML-predicted PPGRs, the primary outcome was difference in start time of PPGR while the secondary outcomes were the ratio of the corresponding 2-hr and 3-hr area under the PPGR curves.
Results: We analyzed data from 19 participants with CGM and self-recorded meal logs (35 ± 6 years, 47% Asian or Black/African-American, 37% Hispanic/Latino). The ML algorithm predicted start time of PPGRs to within a median 30 [19,45] minutes of self-logged meal time (Fig 1). The median ratio of the corresponding 2-hr PPGR AUCs was 1.0 [0.98,1.03] and 3-hr AUCs was 1.0 [1.00,1.01].
Conclusion: An ML algorithm showed promising performance in identifying PPGRs accurately from CGM data in persons with GDM, enabling a new automated approach to meal logging and analyzing postprandial glucose patterns.
S. Barua: None. T. Sangmo: None. A. Khan: None. L. Berube: None. L. Li: None. S. Williams: None. T. Rosen: None. S. Rawal: Stock/Shareholder; Merck & Co., Inc., LabCorp, Pfizer Inc. Research Support; National Institutes of Health. Stock/Shareholder; Bristol-Myers Squibb Company.
Rutgers SHP Dean's Intramural Grant
